A dataset for urban road instance segmentation

本实验使用了一个综合数据集，该数据集结合了 Cityscapes 数据集和 BDD100k 数据集，总共有 5000 张图像。其中，Cityscapes 数据集提供 3475 张图像，而 BDD100k 数据集提供 1525 张图像。由于 Cityscapes 数据集中的大部分图像是在弱光条件下拍摄的，因此在 BDD100k 数据集中专门添加了自然光照条件下的城市场景图像，以增强模型在不同光照条件下的鲁棒性。数据集随机分为 4000 张训练图像和 1000 张验证图像，用于模型训练和性能评估。数据集包含 8 个实例级细分类别，包括行人、骑行者、汽车、卡车、公共汽车、火车、摩托车和自行车。这些类别覆盖了城市场景中常见的目标类型，可以有效支持多类别分割任务中模型的学习和优化。为了便于与现有模型进行比较，将实验数据集转换为 YOLO 和 COCO 格式。这种格式转换不仅提高了数据集的兼容性，而且为评估模型的性能提供了标准化的基准。该数据集的设计旨在提供一个平衡和多样化的训练和验证环境，以支持模型在复杂城市场景中的鲁棒性和准确性。

This experiment employs a comprehensive dataset that combines the Cityscapes and BDD100k datasets, with a total of 5000 images. Specifically, the Cityscapes subset contains 3475 images, while the BDD100k subset contributes 1525 images. Since most images in the Cityscapes dataset are captured under low-light conditions, urban scene images under natural lighting conditions were specifically added to the BDD100k subset to enhance the model's robustness across varying illumination conditions. The dataset is randomly split into 4000 training images and 1000 validation images for model training and performance evaluation. The dataset encompasses 8 instance-level fine-grained categories, including pedestrians, riders, cars, trucks, buses, trains, motorcycles, and bicycles. These categories cover common object types in urban scenarios, effectively supporting the model's learning and optimization in multi-class segmentation tasks. To facilitate comparison with existing models, the experimental dataset was converted to YOLO and COCO formats. This format conversion not only improves the compatibility of the dataset but also provides a standardized benchmark for evaluating model performance. The design of this dataset aims to provide a balanced and diverse training and validation environment to support the model's robustness and accuracy in complex urban scenarios.

搜集汇总

背景与挑战

背景概述

该数据集是一个用于城市道路实例分割的综合数据集，结合了Cityscapes和BDD100k数据集，共包含5000张图像，旨在通过融合不同光照条件（如弱光与自然光）增强模型在复杂城市场景中的鲁棒性。数据集涵盖8个实例级细分类别，包括行人、汽车等常见目标，并已转换为YOLO和COCO格式，以支持多类别分割任务的模型训练和评估，数据随机分为4000张训练图像和1000张验证图像。